Run a software factory from your phone
One person, one Telegram thread and a fleet of agents, all open-source.
I want to send a text like "ship the fix" or "start a research project on X" from my phone and have a system start the work, do the first chunk, and come back with output and the decisions only I can make. My job is to set intent and review results. I don't babysit. Software is the first and hardest use. The same machine can produce documents, plans and research. Tap each layer:
The reference architecture
Every layer is open-source and self-hosted. Code-specific logic lives only in the executor and guardrail layers, so the control plane works for any kind of task.
What stays with me
The machine does the mechanical work: noticing, planning, executing, testing, opening the change. I decide what to build and whether the result is good. That's why the control plane, Hermes, treats a pull request and a research memo the same way. Code-specific parts, the coding agent and the CI guardrail, sit at the edges, so one message can start any kind of project.
Why open-source and any model
Cost: a factory burns tokens. Routing keeps it affordable: mechanical work goes to a cheap or local model, and frontier models only get tasks that need them. So the stack has to point at any provider, including your own GPU. Control: it merges code and runs unattended on your data. Put it on hardware you own, where no vendor can change the terms.
Why Telegram
Factory work is asynchronous. You send a task and get pinged when it's done or needs you. A chat app you already carry fits that better than a dashboard. Send text or a voice note (Telegram transcribes it). When a change needs your eyes, the factory sends a twocents link and you pin what's wrong from the same thread.
A bot that merges code from a phone is powerful, so it only answers to my Telegram ID. Approving from a phone also means less context, which is why the auto-merge guardrails matter so much.
What I'm building
I'm building it twice as a bake-off: the same factory with two coding
executors, oh-my-pi and OpenCode, under Hermes. I'll publish a
guide per layer as I set it up, with cost per task, edit reliability and how much each tool
fights me. I only publish what has worked in my setup. New to this? Start here: